python src/prophet_v1.py
- Python 3.7.9
- conda
- library(requirements.txt)
- Ubuntu 20.04
- CPU: i9 9900K
- Memory: 16 * 2 (GiB)
Since there is no data for days when there was no trade, I have added such data that such days would also be zero. ( Adding it increased the score compared to not adding it.)
I didn't do features engineering just to prepare the data.
I did the modeling based on prophet.
I made a "logistic" prediction because the corona can cause the trend to be negative. See the parameters for details.
The discussion I referred to ↓ facebook/prophet#1668
params = {
"growth": "logistic",
"changepoint_prior_scale": 0.05,
"seasonality_prior_scale": 30,
"mcmc_samples": 0,
"seasonality_mode": "multiplicative",
}
Holiday data were not included because they were a factor in over-fitting.
- LGBM modeling(ex. https://www.kaggle.com/kneroma/m5-first-public-notebook-under-0-50)
- complex parameters
- features(on Prophet)